To Improve Protein Sequence Profile Prediction through Image Captioning on Pairwise Residue Distance Map

Sheng Chen, Zhe Sun, Lihua Lin, Zifeng Liu, Xun Liu, Yutian Chong, Yutong Lu, Huiying Zhao, Yuedong Yang

TL;DR

This paper introduces SPROF, a novel deep learning method that leverages 2D distance maps of protein structures, inspired by image captioning, to predict protein sequence profiles. SPROF significantly outperforms previous methods by achieving a 5.2% improvement in sequence recovery, demonstrating the power of using structural 3D information encoded in distance maps for this task.

What Problem Does This Paper Address?

Predicting protein sequence profiles is crucial for advancing protein design, allowing researchers to generate multiple sequences that can fold into a desired structure and perform a specific function. Traditional methods often rely on 1D structural properties, which are insufficient to capture the complex 3D nature of proteins. While some methods have integrated neural networks to improve predictions, they still struggle to fully utilize the rich information present in a protein's 3D structure. This paper aims to bridge this gap by developing a method that effectively incorporates 3D structural information into sequence profile prediction.

Methods

The core innovation of this work is the development of SPROF, a deep learning architecture inspired by image captioning. Instead of relying solely on 1D structural features, SPROF utilizes a 2D distance map, which represents the pairwise distances between amino acid residues in a protein's 3D structure. This distance map is treated as an image, and the prediction of a sequence profile is framed as generating a "caption" for this image.

SPROF's architecture combines:

  • A Residual CNN (ResNet) with Self-Attention: This module processes the 2D distance map to extract relevant features, focusing on important regions through a self-attention mechanism.
  • Bidirectional Long Short-Term Memory (BiLSTM) RNN: The features extracted from the CNN, along with the original 1D structural features, are fed into a BiLSTM. This allows the model to learn long-range dependencies within the protein sequence.
  • Fully Connected Layers: These layers process the output of the BiLSTM to generate the final protein sequence profile, predicting the probability of each of the 20 amino acid types at each position.

The paper highlights that this is the first method to employ a 2D distance map for predicting protein properties in this manner.

Datasets

The study utilized a dataset curated in 2017, consisting of 12450 non-redundant protein chains with specific resolution, R-factor, and sequence length criteria. From this, 11200 chains were selected for training and 1250 for an independent test set. Due to GPU memory limitations, chains with 500 or more amino acids were removed, resulting in a final training set of 7134 chains (TR7134) and a test set of 922 chains (TS922).

Key Insights and Ideas

  • 2D Distance Maps as a Rich Source of 3D Information: The success of SPROF demonstrates that 2D distance maps, derived from 3D structures, contain crucial information for predicting protein sequence profiles, outperforming methods that rely solely on 1D features.
  • Image Captioning Analogy is Effective: Framing protein sequence profile prediction as an image captioning task, where the distance map is the image and the sequence profile is the caption, proves to be a powerful paradigm.
  • Self-Attention Enhances Feature Extraction: The inclusion of a self-attention mechanism in the CNN module allows the model to selectively focus on the most informative parts of the distance map, contributing to improved accuracy.
  • Long-Range Interactions are Key: The BiLSTM component effectively captures long-range dependencies between residues, which are critical for understanding protein structure and function.
  • Binding Sites vs. Non-Binding Sites: The study found a significant difference in prediction accuracy between residues in binding sites and non-binding sites, suggesting that this method could be useful for identifying functional regions.

Key Results

  • Significant Improvement over SPIN2: SPROF achieved a sequence recovery rate of 39.8% on the independent test set, a 5.2% improvement over the previous state-of-the-art method, SPIN2 (34.6%).
  • Robustness Across Metrics: This improvement was consistent across various evaluation metrics, including top-1 and top-2 matches, precision, and recall scores, and across different protein lengths and test sets.
  • Feature Importance: The 2D distance map was identified as the most crucial feature, contributing a 6.6% increase in sequence recovery when excluded. Energy-based features also showed a notable contribution.
  • Case Study Success: In a case study of the precorrin-6A reductase cobK protein, SPROF achieved 60% accuracy for a specific fragment, significantly outperforming SPIN2's 26.7%. SPROF correctly predicted 11 amino acids that SPIN2 missed, particularly hydrophobic residues.

Limitations and Future Work

  • GPU Memory Constraints: The removal of longer protein chains due to GPU memory limitations suggests that future work could explore more memory-efficient architectures or hardware.
  • Potential for Other Applications: The authors suggest that the network architecture, which effectively utilizes 2D distance maps, is expected to be beneficial for other 3D structure-based applications, such as binding site prediction, protein function prediction, and protein interaction prediction.
  • Discrimination of Functional Residues: The observed difference in prediction accuracy between binding and non-binding residues indicates that SPROF's output could be used to discriminate functional residues.
  • Advancing Protein Design: The improved prediction of sequence profiles could further advance existing protein design and fold recognition techniques.

Summary updated

Abstract

ABSTRACT Protein sequence profile prediction aims to generate multiple sequences from structural information to advance the protein design. Protein sequence profile can be computationally predicted by energy-based method or fragment-based methods. By integrating these methods with neural networks, our previous method, SPIN2 has achieved a sequence recovery rate of 34%. However, SPIN2 employed only one dimensional (1D) structural properties that are not sufficient to represent 3D structures. In this study, we represented 3D structures by 2D maps of pairwise residue distances. and developed a new method (SPROF) to predict protein sequence profile based on an image captioning learning frame. To our best knowledge, this is the first method to employ 2D distance map for predicting protein properties. SPROF achieved 39.8% in sequence recovery of residues on the independent test set, representing a 5.2% improvement over SPIN2. We also found the sequence recovery increased with the number of their neighbored residues in 3D structural space, indicating that our method can effectively learn long range information from the 2D distance map. Thus, such network architecture using 2D distance map is expected to be useful for other 3D structure-based applications, such as binding site prediction, protein function prediction, and protein interaction prediction.
BibTeX
@article{chen2020improve,
  author = {Sheng Chen and Zhe Sun and Lihua Lin and Zifeng Liu and Xun Liu and Yutian Chong and Yutong Lu and Huiying Zhao and Yuedong Yang},
  title = {{To Improve Protein Sequence Profile Prediction through Image Captioning on Pairwise Residue Distance Map}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2020},
  doi = {10.1021/acs.jcim.9b00438},
}

Loading related content...

Figure